Compare artificial and biological neural networks. insist that computers do not learn, and are only

3. Compare artificial and biological neural networks. insist that computers do not learn, and are only

What aspects of biological networks are not mim- taught by humans. Do you agree? Please comment.

icked by artificial ones? What aspects are similar?

. C H A P T E R 1 2 A D V A N C E D INTELLIGENT SYSTEMS

combined cffects of the summation and transforma- the "no" output node. Discuss what may have hap- tion functions, and how they differ from statistical

pened and whether the applicant is a good credit risk. regression analysis.

9. Everyone would like to make a great deal of money 5. A N N s can be used for both supervised and unsuper-

in the stock market. Only a few are very successful. vised learning. Explain how they learn in a supervised

Why is an A N N a promising approach? What can it mode and in an unsupervised mode.

do that other decision-support technologies cannot 6. Review the A N N development process. Compare it

do? H o w could it fail?

to the development process of expert systems. U s e an 10. Similarly, describe how an investor can use genetic example to show both processes, and list their similar-

algorithms to make a fortune on the stock market? ities and differences.

Can a GA-based system perform better than an N N - 7. Explain the difference between a training set and a

based system? Why or why not? testing set. Why do we need to differentiate them?

11. Describe three advantages of fuzzy reasoning and Can the same set be used for both purposes? Why or

provide an example to support each. If you disagree why not?

in any of the three cases, explain why. 8. Real-world scenario: A neural network has been con-

12. Describe the advantages and disadvantages of inte- structed io predict the creditworthiness of applicants.

grating multiple methods for developing intelligent There are two output nodes, one for yes (1, yes; 0, no)

systems. Describe all the possible integrations and one for no (1, no; 0, yes). An applicant received a

between CBR, A N N , G A , fuzzy logic, and rule-based score of 0.83 for the "yes" output node and a 0.44 for

systems, and assess their feasibility.

• EXERCISES 1. Access the University of Kaiserslauten's case-based

reasoning Web site (www.agr.informatik.uni-kl.de/

w,

Isa/cbr/cbr-homepage.html). Examine the latest CBR research and demo software. How is CBR different from rule-based concepts? Try some reasoning soft-

ware, compare the method to rule-based inferencing, and write up your experience in a report.

2. Identify a news group that is interested in case-based

\j\k>A

reasoning. Post a question regarding recent successful applications and see what feedback you get. What are the latest concerns and questions?

3. For each of the following applications, would it be better to use a neural network or an expert system? Explain your answers, including possible exceptions

c. Compute the value of Z with the sigmoid transfer or special-conditions.

function used at all neurons. a. Diagnosis of a well-established but complex dis-

5. Using the Braincel neural network package (down- ease

load a demo from the Web, www.promland.com) or b. Price-lookup subsystem for a high-volume mer-

another, build, train, and test a neural network to chandise sale

solve the following simple regression formula that c. Automated voice-inquiry processing system

predicts Y as a function of X (for this simple case, d. Training new employees

train the network on the entire set of data): e. Handwriting recognition

XY

4. Review this neural network and compute Z.

where X, = 15, X, = 8, X, = 14, W 1 = 0.6, W 2 = 0.3,

W 3 = 0.1, weight for 7, = 0.6, weight for

a. Compute the value of Z without a transfer func-

b. Compute the value of Z with a threshold func-

P A R T I V INTELLIGENT DECISION SUPPORT SYSTEMS

In a spreadsheet, plot the data, and solve the simple The weights from the hidden layer neurons to the linear regression formula Y = a + bX. Compare the

output node are

estimates made by the neural network and by the 0.3,0.4,0.2,0.2. regression. Calculate the sum-of-the-squares error

7. Express the following statements in terms of fuzzy and R 2 for each. Which one is better?

sets:

a. The chance for rain today is 80 percent. (Rain? No credit approval system. There are three neurons in

6. Develop a deployable neural network code for a

rain?)

the input layer, four in the hidden layer, and one in

b. Mr. Smith is 60 years old. (Young?) the output layer. Three input factors were considered

c. The salary of the President of the United States is for a potential client: credit rating, debt to income

$250,000 per year. (Low? High? Very high?) ratio, and net income. They have already been scaled

d. The latest survey of economists indicates that they so that the lowest value for any potential customer is

believe that the recession will bottom out in April

0 and the highest value is 1. These are the inputs. The (20 percent), in May (30 percent), or in June (22 output represents credit approval. If the output is 1,

percent).

8. Compare the effectiveness of genetic algorithms ver- is denied. The transfer function is a threshold func-

then credit is approved. If the output is 0, then credit

sus standard methods for problem solving as tion. If the total flow arriving in a hidden layer neu-

described in the literature. How effective are genetic ron is 0.5 (one-half) or greater, then, the flow out is

algorithms?

set to 1, otherwise it is set to O.The weights from the

9. Investigate custom online newspaper systems on the three inputs to the hidden layer are (rows = from

Web (e.g., CRAYON) and evaluate how effective inputs, columns = to hidden nodes)

they are at filtering the news so that you get only the material you really want. Which ones worked

Hidden Node

best?